Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions NEXT_TASKS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Edge-review next tasks

**Prerequisite repairs verified on M5, 2026-11-17; unmerged [PR #211](https://github.com/CultureBotAI/EdgeMech/pull/211):** two fresh environments at `3057bb2` passed actual native canaries, canonical save/readback and 380 frozen input bindings each; [live results](docs/validation/m5-prerequisites-live-20261117.md) retain the first failed attempt and #212 bootstrap repair. The corrected first clone passed before the second was provisioned. Exact-runtime CI passes 1,352 tests, 16 contracts, Ruff and builds. #203 binds future instructions/package/session paths; #205 clarifies fidelity proof dependencies; #206 makes setup reproducible; #207/#208 prevent unsafe staging/drift; #210 preserves substantive definition text with bounded omissions. No scientific round started. Next: recover missing original source/identity/definition witnesses using the ten-case map, beginning with bounded MetaTraits mapping/polarity replay, and finish typed adjudication inputs (#209). Retained-90/9,900 stay closed; old rounds and failed attempts remain unchanged.

**Round03 finalized failed, 2026-11-16:** [Round03 results](docs/validation/m5-calibration-round03-results-20261116.md) preserve two fresh sealed matrices and separate adjudication. Initial agreement is 54/60, but source-fidelity agreement 7/10 and final resolution 3/10 fail their unchanged 8/10 floors. Six initial claim/model judgments and three supplemental source-fidelity judgments are unsupported; four disagreements remain unresolved, including one blocking model judgment. Mechanical finalization passes 20/20 canonical pairs, 153/153 selected initial proofs and 55/55 final citation rows, with 59 final local proofs separately validated. Eleven rejected authoring attempts remain preserved. A missed historical producer clause independently fails blinding (#200); the earlier blanket preparation audit is corrected. All scientific workers exited with ending bindings intact. Remaining90/9900 remain closed. The next scientific step requires missing original source/identity/definition witnesses and a separately identified round with unchanged thresholds.

**Engineering and delivery, [PR #190](https://github.com/CultureBotAI/EdgeMech/pull/190):** publication #199, future producer views #200, preseal frozen kind/version #201, report delivery #202 and publication import setup #204 are repaired and independently reviewed. The first failed derivative canary remains preserved. New runtime `2e31b6f` passed 1,290 tests, 192 focused tests, 16 contracts, Ruff, builds, exact CI, a fresh same-clone production save/check, an inspected derivative canary and full twenty-review publication replay. The [bounded result bundle](artifacts/edge-calibration-round03-20261116/README.md) preserves the failed scientific result, all six confusion tables and original accounting. Future science remains blocked by missing original witnesses and deferred instruction/package coherence [#203](https://github.com/CultureBotAI/EdgeMech/issues/203). Shared CLAW `4fad8df` remains reviewed but unmerged. The directional exercise uses a primary abstract; full-paper inspection is not claimed. Old rounds, inputs/seals, failed attempts and owner candidates stay intact. No new run may reuse a receipt after its runtime bindings change.
Expand All @@ -10,7 +12,7 @@ The earlier #152 operational save/check canary passed and remains excluded from

**Completed round01, 2026-11-16, [PR #150](https://github.com/CultureBotAI/EdgeMech/pull/150):** [Round01 failed after execution and separate adjudication](docs/validation/m5-calibration-20261116.md). Two fresh same-model sessions sealed 10-by-6 matrices; initial agreement 56/60, but object-node agreement 7/10 fails the required 8. Independent audit found eight unsupported initial modeling labels, source-fidelity resolution 7/10 and two blocking disagreements. All 18 available canonical pairs replayed, but both fresh-05 saves failed; all saved reviews remain partial. Original matrices and failed saves remain unchanged. Issues #152–#157 track acceptance semantics, predicate reasoning, METPO hierarchy, local definitions, fidelity/exposure and aggregate quotation budgets; #145 tracks fresh-04 identity. Helper findings #144/#148/#149/#158/#159 are repaired. Final checks: 830 tests, full Ruff scope, contracts and package builds passed. [Round02 registration and revised rubric](docs/plans/m5-calibration-round02-20261116.md) preserve every original threshold. Remaining-90 and 9,900 gates stay closed. Both original M5 workers have exited; the release worker delivered the source-only candidate referenced above. Current follow-up execution is reported separately in the round02 entry.

Last reconciled: **2026-11-16** (America/Los_Angeles), against `main` at `f841b1dff874e78d1a14075413be6a92c90f1378`, merged PRs #135/#131/#128/#120 and open PRs #102/#37. PR #131 is merged and #127 is closed. #130's bounded source quarantine has been independently verified against regenerated producer bytes and merged in [KG-Microbe PR #1867](https://github.com/Knowledge-Graph-Hub/kg-microbe/pull/1867); graph-release verification remains pending. Independent scientific calibration has not passed.
Last reconciled: **2026-11-17** (America/Los_Angeles), against `main` at `f841b1dff874e78d1a14075413be6a92c90f1378`, merged PRs #135/#131/#128/#120 and open PRs #211/#190/#167/#151/#150/#136/#102/#37. Shared CLAW PR #1289 remains open at `4fad8df`. PR #131 is merged and #127 is closed. #130's bounded source quarantine has been independently verified against regenerated producer bytes and merged in [KG-Microbe PR #1867](https://github.com/Knowledge-Graph-Hub/kg-microbe/pull/1867); graph-release verification remains pending. Independent scientific calibration has not passed.

Scope: the stratified edge review and related edge-quality follow-ups. Update this file when work starts or ships, checking GitHub and current code before changing status. Record completion dates and PRs. Preserve historical manifests, packets, reports and scientific decisions; a merged engineering PR does not promote an edge's scientific status. Unrelated release/deployment work is outside this backlog.

Expand Down Expand Up @@ -38,7 +40,7 @@ Evidence: [fresh-canary report](artifacts/edge-recanary-20261113/RECANARY-REPORT
| 2 — **PENDING, actionable bounded audit** | Semantic identity, relation and node-category correctness | Prepare pinned authorities and audit the selected calibration cases, then the existing 100 after gates pass. Prioritize [#68](https://github.com/CultureBotAI/EdgeMech/issues/68)/[#106](https://github.com/CultureBotAI/EdgeMech/issues/106) identity joins and [#75](https://github.com/CultureBotAI/EdgeMech/issues/75)/[#86](https://github.com/CultureBotAI/EdgeMech/issues/86)/[#87](https://github.com/CultureBotAI/EdgeMech/issues/87) relation/type-strain/bin semantics. Review nonempty categories, not only missing fields. All 22 fresh-canary biological-literature exclusions still receive identity/modeling review. |
| 3 — **PENDING, actionable retained-evidence work** | DOI identity, exact snippets and categorical claim judgments | Resolve publication identity [#69](https://github.com/CultureBotAI/EdgeMech/issues/69); replay retained evidence for the selected calibration cases first, then retrieve missing relevant text through appropriate channels. Verify literal quotes and independently judge entailment, conditions and organism/strain scope. Record support, refutation, conflicts, insufficient evidence and assessment blockers distinctly. |
| Parallel — **IN PROGRESS** | General full-text credential protection: [#100](https://github.com/CultureBotAI/EdgeMech/issues/100), [PR #102](https://github.com/CultureBotAI/EdgeMech/pull/102) | The draft has a clean independent review and green CI at `2364161819951b7899ba62cebcd461bf0394c4a7`. Reconcile with current main and validate before a separately authorized merge. Complete this before further authenticated general-helper retrieval; it need not block local modeling or retained-evidence review. |
| 4 — **ROUND02 FAILED AFTER ADJUDICATION** | Preserve results and repair prerequisites before a separate round | Initial agreement 52/60, source-fidelity agreement 4/10, final source-fidelity resolution 7/10 and evidence integrity fail. All 20 canonical slots replayed as missing; audit covers every emitted initial/final resolved claim/model label and all 60 final axes. Preserve three unresolved source-fidelity disagreements, failed proof attempts and exposure #175; repair clone-origin preparation under #176 and exact citation/proof bindings before new canaries and a separately frozen round. Retained-90 and 9,900 review remain closed. |
| 4 — **ROUND03 FAILED; engineering prerequisites verified in unmerged PR #211** | Recover original witnesses before a separate round | Round03 achieved 20 canonical pairs, but source-fidelity agreement 7/10, resolution 3/10, unsupported initial labels and blinding fail the unchanged gates. PR #211's repaired runtime passes two actual same-clone canaries/save/readback checks after preserving the initial #212 failure. Next recover source selection/identity/field-meaning witnesses in the ten-case map and complete typed adjudication inputs #209; then freeze a new scientifically reviewed package and run its own readiness checks. All historical failures stay intact. Retained-90 and 9,900 remain closed. |
| 5 — **PENDING; upstream fixes/releases where needed** | Correct and verify affected assertions and nodes | File each finding, identify its maintained owner, retain before/after proposals, add regressions, regenerate a new candidate and independently re-assess affected occurrences. Include [#56](https://github.com/CultureBotAI/EdgeMech/issues/56), [#70](https://github.com/CultureBotAI/EdgeMech/issues/70), [#72](https://github.com/CultureBotAI/EdgeMech/issues/72) and proof of existing fixes [#71](https://github.com/CultureBotAI/EdgeMech/issues/71)/[#73](https://github.com/CultureBotAI/EdgeMech/issues/73). A merged source PR is distinct from a verified corrected graph release. |

**Completed initial execution, 2026-11-16:** M5 main was synchronized to `f841b1d` for the original [execution plan](docs/plans/m5-review-execution-20261116.md). The original calibration and BactoTraits verification workers ran and exited. Round01 failed; the BactoTraits-only candidate passed its bounded verification, while full canonical release/publication remains outstanding. Round02 subsequently failed initial agreement, final source-fidelity resolution and evidence integrity after separate adjudication; both initial reviewer jobs exited. Preserve sampled/current authority distinctions, all original thresholds and both rounds' immutable failures. Retained-90 and 9,900 review remain gated; neither operational startup nor source-only candidate verification establishes scientific calibration or full-release completion. This reconciliation retains #165's historical execution context.
Expand Down
90 changes: 90 additions & 0 deletions docs/plans/m5-prerequisite-package-20261117.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
# M5 prerequisite package and actual-environment checks

This engineering procedure addresses #203, #205 and #206, with independent
review fixes #207 and #208. It does not start another calibration or replace
the failed round03 result. The scientific invocation contract additionally
rejects adjudication until its initial matrices, seals and comparison have a
reviewed typed binding (#209).

## Inputs and boundaries

Use `prepare_m5_prerequisite_package_20261117.py` in a freshly provisioned clone
at the exact committed candidate runtime. The retained source is M5's
`m5-calibration-round03-20261116` worktree. The repaired engineering manifest is
`.temp/m5-round03-worker-20261116/publication-runtime-02/manifest.json`, SHA256
`1e0ad8f7dea88d5138d891341d46ac772f8a7896f0282ee11ddaf2b0686fd168`.
Its conservative operational specification is
`.temp/m5-round03-worker-20261116/engineering-operational-canary-spec-02.json`,
SHA256 `3321872d3bb7685836b46d0d799973ffb17335ea6b78a5783bf63c941a7c1e32`.
Also supply `--bootstrap-receipt-sha256
894a803fc18fdf66e715884f51a899fdb5ba5107cec3ad13b237c9c9e369f7af`.
This binds the original `fresh-47/receipt.json` outside the retained manifest.
Its original timestamp and bytes stay unchanged; it is not a new runtime/save
receipt. Configuration equality and the maintained preparation checker are
required before it can be reused for byte replay.

The helper checks both exact input hashes before staging. Retained evidence,
source packets, original records and bounded-view receipts retain their exact
bytes and case/evidence identities. Current tracked runtime files receive new
pins, with old/new identities recorded in a transition ledger. Historical
scientific instructions are excluded from instruction roles. The only active
instruction is the operational replay canary. New v4 case views have distinct
paths; their inclusion does not itself establish blinding or scientific readiness.

Every maintained execution input, including hidden/ignored executable helpers,
is pinned using the launcher's inventory. The original cohort, protocol,
thresholds, schema and shared contract remain bound. The new neutral procedure,
coordinator witness map and setup audit may be supplied with repeated
`--prerequisite-document`; they receive coordinator-only visibility here.

All destinations must be new and private under Git-ignored `.temp`. Metadata,
documents, canonical review destination and source pins are checked before
copying. The helper replays planned evidence hashes, document pins, execution
inventory, HEAD and clean tracked state around freezing and authoring. It
preserves failed attempts rather than overwriting them or emitting success.
The bootstrap quartet is copied at mode 0600 inside a 0700 case directory, as
required by the real checker; other sealed package inputs retain their existing
mode policy. The actual free `prepare_calibration_20261115.py --check fresh-47`
must pass before freezing succeeds. It also reads the retained private source
archive on M5, verifies its expected upstream hashes, and leaves it unchanged.
The public fresh-canary projection is explicitly pinned. This addition fixes
the first actual native-worker failure (#212), which remains preserved.

## Execution sequence

1. Provision exactly one fresh M5 reviewer clone with the maintained setup
helper. Inspect its on-disk receipt, origin, revision and isolated environment.
2. Run package preparation with `--plan`, then the identical arguments and
`--run`. Supply the source root/manifest/hash and operational spec/hash above;
the bootstrap receipt hash, a new `--output`, `--package-id`, UTC `--timestamp` and matching timestamped
`--review-id` are required. Inspect the persisted preparation receipt.
3. Use the maintained worker launcher with role `canary`, the new manifest,
`operational/review.json`, `operational/proofs.json`, the original fresh-47
packet, and `scripts/adhoc/m5_review_20261116/replay-canary.md`. First plan,
then run through the actual native Codex launcher with a new output root.
4. Inspect the actual canonical YAML/Markdown pair, its exact hashes and the
parent production save/check validators. Confirm the receipt identifies this
clone, runtime and manifest. A child success message alone is insufficient.
5. Only after that whole unit passes, provision and check a second fresh clone
through the same path. A changed runtime or failed canary requires a new
verified canary before any further fanout.

The operational fixture stays `scientific_review: false`, `not_assessed`, with
all six axes unresolved. It is outside scientific matrices, canonical scientific
pair denominators and graph admission. A new engineering manifest explicitly
uses `package_purpose: engineering_prerequisites`; the scientific invocation
contract rejects it even when all production checks pass.

## What remains scientific work

This freezes complete **available byte bindings**, not missing original
witnesses. The coordinator witness map names the unresolved source selection,
identity, field meaning, units, bin interpretation and version dependencies.
The bounded MetaTraits mapping/polarity replay is a concrete next recovery task.
No new literature service, primary full-paper inspection, scientific rating,
graph release or expansion is part of these operational checks. Retained-90
and 9,900 remain gated on an independently evaluated successful calibration.

Live results belong in a separate validation record naming the actual runtime
and receipts. A later documentation commit does not retroactively change a
frozen clone's runtime or make its receipt valid for another commit.
Loading
Loading